Train a Flux LoRA: Complete Walkthrough (Free and Paid Routes)

To train a Flux LoRA, pick one of three routes: FluxGym on your own NVIDIA card with 12 GB of VRAM or more, Civitai's on-site trainer paid in Buzz, or a hosted trainer on fal or Replicate. Each one takes a set of captioned images and a trigger word and hands back a LoRA file. A default FLUX.1 run is listed at about $1.46 on Replicate, $2 on fal and 2,000 Buzz on Civitai, and FluxGym costs only your GPU time.

Settings and prices checked in October 2026 against the FluxGym README and source, AI Toolkit's FLUX.1 example config, Civitai's developer docs for Flux 1 and Flux 2 Klein, its trainer guide and pricing page, fal's FLUX.1 fast trainer, Replicate's fine-tuning guide and Black Forest Labs' klein training guide. Every setting below is that trainer's documented default or example. We did not benchmark the routes against each other, so nothing here says which output looks best.

A LoRA is one of several ways to hold a face. If you have not settled on training yet, the consistent character AI guide compares it with reference-image methods first, and Flux consistent character covers the no-training route on Flux itself.

What you will have at the end

One LoRA file per saved checkpoint, a trigger word that switches the character on, and a short test that tells you which checkpoint to keep. The job has five stages, and only the fourth differs between the three routes.

Flux LoRA training guide: choose the Flux model first

Train on the model you will generate with. "Flux" now covers several architectures, and a LoRA does not move between them.

Licences are from Black Forest Labs' FLUX [dev] Non-Commercial License and its klein training guide. The dev licence allows commercial use of outputs but limits the model itself to non-commercial use, so read it before you train a persona you earn from. FLUX 3, the newest family, runs through the API only: the release notes list open weights as a later phase, so there is nothing to train a LoRA on yet.

Three ways to train a Flux LoRA, compared

The routes differ in what you need to own, which Flux models they cover and how you pay.

• Pick FluxGym if you own the card and want unlimited retries on FLUX.1.
• Pick Civitai if the persona is fictional, you want no setup, and you want Klein 4B at the lowest listed price.
• Pick fal or Replicate if you want a file in minutes, or need the run inside your own code.

Flux character LoRA: the dataset comes before the trainer

Every route takes the same input, and it decides more of the result than any setting. Vendor docs put a face at 10 to 20 images and a full character at 20 to 40; how many images to train a face LoRA lists each source, and the character LoRA dataset checklist gives a shot list.

Caption in plain sentences. Civitai's trainer guide calls natural-language captions the preferred method for Flux, and Replicate's guide says to choose a trigger word that is not an existing word and to avoid TOK, which clashes with other fine-tunes.

Route 1: train a Flux LoRA free with FluxGym

FluxGym is a free web interface over kohya sd-scripts that trains FLUX.1 LoRAs on 12, 16 and 20 GB cards. The software costs nothing; the GPU is yours. Its repository was last updated in July 2026.

Look at two defaults before the first run. FluxGym resizes images to 512 and ships with a learning rate of 8e-4 and a rank of 4, both under the advanced accordion. That rate sits above the 5e-4 Civitai's Flux 1 docs say to stay under, so if samples burn early it is the first number to lower. Face LoRA training steps and learning rate covers the documented ranges, and the free LoRA Training Steps Calculator reproduces FluxGym's step arithmetic and adds a time and GPU-cost estimate.

No suitable card? Renting one to run FluxGym works, but it turns the free route into a paid one. Train a LoRA free online compares the free quotas that exist instead.

Route 2: Civitai's on-site trainer

Civitai trains Flux in the browser and bills in Buzz. Two limits come first. Its own notice says the site has not been accessible in Australia since March 16, 2026, and its rules prohibit content based on any real person's likeness, so this route is for fictional personas only.

1. Start the wizard. Click Create, then Train a model, choose a model type and name it.
2. Add the images. Upload a zip of images with matching .txt captions, or loose images, and tick the acknowledgements.
3. Caption. Auto Label with the Caption type writes natural-language captions, and Prepend Tags puts your trigger word on every one.
4. Choose the base model. Flux.1 Dev, Flux.2 Dev, or Flux2 Klein 4B or 9B Base. Result: the Buzz cost appears before you submit.
5. Add sample prompts and submit. Result: an email when the run finishes, and each epoch listed with its samples.
6. Download within 30 days. Civitai keeps at most 20 epochs and deletes unclaimed files after that.

The developer docs give the defaults: 2,000 steps, 10 saved epochs, a 1e-4 learning rate with adamw8bit on a cosine schedule, and a rank of 16 on Flux 1 or 32 on Klein. A run that fails is retried, then refunded in full if it fails completely. Our Civitai LoRA training walkthrough goes through both of its trainers and every setting.

Route 3: a hosted trainer on fal or Replicate

Both services run a FLUX.1 fast trainer from a web form or an API call. You upload a zip, set a trigger word and wait a few minutes.

1. Zip the dataset folder. Include the caption files if you wrote them.
2. Open the trainer. On fal, FLUX.1 fast training. On Replicate, the fast-flux-trainer form from its guide.
3. Set the trigger word and, on Replicate, set lora_type to subject.
4. Leave steps at 1,000 for the first run. Replicate's guide says fewer under-learns and more adds cost without much gain.
5. Start it. Result: Replicate quotes about two minutes for 20 images; fal says minutes, not hours.
6. Generate with the trigger word, then download the weights if you want the file elsewhere.

fal also lists a FLUX.2 [dev] trainer and a klein 9B trainer. Its klein 4B trainer page is marked as no longer supported (checked October 2026), so for Klein 4B use Civitai or AI Toolkit.

Flux LoRA settings: each trainer's documented defaults

These are the numbers each trainer ships with. Start from the row for your trainer, not from a forum recipe.

For a settings file to copy, AI Toolkit publishes one. This excerpt holds the values that matter from its train_lora_flux_24gb.yaml; copy the full file from the repository, set the folder path and trigger word, and run it with python run.py. A comment in the file calls 500 to 4,000 steps a good range.

network:
  type: "lora"
  linear: 16
  linear_alpha: 16
save:
  save_every: 250
  max_step_saves_to_keep: 4
datasets:
  - folder_path: "/path/to/images/folder"
    caption_ext: "txt"
    caption_dropout_rate: 0.05
    resolution: [ 512, 768, 1024 ]
train:
  batch_size: 1
  steps: 2000
  optimizer: "adamw8bit"
  lr: 1e-4
model:
  name_or_path: "black-forest-labs/FLUX.1-dev"
  is_flux: true
  quantize: true

For Klein, Black Forest Labs' guide gives a learning rate of 8e-5 to 1e-4 and 800 to 1,200 steps for a character LoRA of 10 to 15 images, and it lists 12 GB of VRAM as the minimum for 4B and 22 GB for 9B.

What a Flux LoRA training run costs

Civitai prices a run in Buzz, and its training page gives the conversion as 1,000 Buzz to $1. Its Buzz guide says every type of Buzz, including the free Blue Buzz you earn on the site, can pay for training, so the cash cost can be zero if you have earned enough. The price is steps times a per-step cost plus a charge per saved epoch, and it never drops below 80 percent of the default, so cutting a Flux 1 run to 1,000 steps still costs 1,600 Buzz.

Budget for several runs, not one. A first character LoRA usually takes a few attempts, which is the real argument for a local card if you already own one.

Whose face each route lets you train

Sources: Civitai's real-people rules, fal's acceptable use policy, Replicate's acceptable use policy and Black Forest Labs' usage policy, checked October 2026. The practical rule: train an original AI persona or your own face. Anyone else needs written consent, and even then not on Civitai.

Test the checkpoints, then keep one

The last file is not automatically the best one. Run the same three or four prompts, with the same seed, against every saved checkpoint: a neutral portrait, a profile, a full-body shot and an outfit that is not in the dataset. Keep the checkpoint that holds the face and still obeys the prompt. The LoRA training guide walks through that comparison.

A trained file is the start of a persona, not the end. The AI Influencers program covers the build in order, from creating the persona through consistent faces with LoRA training to body consistency with ControlNet and the video lessons after it.

Troubleshooting a Flux LoRA run

If the samples keep repeating one pose or one background, work through the symptom chart in face LoRA overfitting before you change a setting.

Training a Flux LoRA: FAQ